{"id":"7eecd19c-888d-4238-bebf-76ad7fbc97be","arxiv_id":"2505.09833","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Rock pushability on rough terrain is predicted from rock volume, shape, and ground slope together with push force feedback using a Bayesian linear regressor.","lead":"This paper proposes a robot system that tests whether rocks can be pushed aside on rough terrain, using camera data and force feedback instead of always driving around them. The goal is to let planetary rovers and other repeated-traffic robots clear obstacles so future trips become faster.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Peak-force threshold is never validated against actual rock movement, so the central pushability classification is unsupported.","rationale":"The paper's intended contribution is a decision system: predict f_max and threshold it against robot capability δ to output pushable or not. The single load-bearing condition is that f_max is a valid observable for the latent property 'rock moves when pushed.' This condition is neither established theoretically nor empirically. The problem definition in Section III-A equates pushability with f_max conditioned on visual features, but the data collection procedure records only force feedback; no post-push displacement or binary movement label is stored. The 'empirically observed' 20 N statement in Section IV is anecdotal and cannot substitute for labels, especially because Section V-B itself lists multiple possible movement outcomes depending on contact pose, center of mass, and terrain. Consequently, the trained model only learns to predict a force value under the experimental push protocol; it could be internally consistent and still fail to indicate pushability if, for example, a rock begins moving only after a peak force spike, or an immovable rock produces a low peak because the arm slips. The decision rule's threshold δ is also under-specified: Algorithm 1 uses robot capability, while the text uses both δ and an empirical 20 N value, and no procedure is given for setting δ from measured force limits. The train/test leakage identified by the reader is a serious additional problem for any predictive claim, but even a perfectly split dataset of force labels cannot validate the affordance classification without movement ground truth. The proposed concrete test directly supplies that missing ground truth by measuring actual rock displacement. Because this gap is foundational to the central claim, the reader's REJECT verdict is appropriate and no change to the verdict is needed.","tokens_in":10670,"tokens_out":3620,"duration_ms":40540,"concrete_test":"Collect a movement-labeled dataset from the 90 simulation pushes: before and after each push, record the rock pose from the point cloud and label the run as 'moved' if centroid displacement exceeds a threshold (e.g., 5 cm). Train the same BayesianRidge model on features (V, E, N) from a subset of runs and evaluate Algorithm 1's decision (f̂_max ≤ δ) on held-out runs against the movement labels, reporting a confusion matrix and balanced accuracy. If rocks that moved have predicted f_max > δ, or rocks that did not move have predicted f_max ≤ δ, in substantial numbers, the peak-force proxy is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the Bayesian model predicts the peak force f_max and Algorithm 1 classifies a rock as pushable when the predicted peak is within the robot's capabilities δ. But pushability is explicitly defined as the probability that the rock moves (Section III-A), and movement is never measured as ground truth. The dataset labels every timestep of a push run with the maximum force magnitude recorded during that run; no binary or continuous movement outcome is captured. The decision rule therefore evaluates a proxy, not the affordance itself. Section V-B even enumerates possible contact outcomes—translational movement, rotational movement, slipping, or no movement—and notes these depend on contact pose and local surface geometry. Without movement labels, an identical force trace from a stalled arm and from a rock that moved after static friction was overcome cannot be distinguished, so the mapping from f_max to pushability is unverified. The real-world experiments only validate VPP volume and surface-normal estimation, not pushability decisions. The simulation section asserts an empirical 20 N threshold but does not derive it from measured movement outcomes, and δ itself is not specified as an operational value. Thus the core classification claim is not supported by the reported evidence even before considering the train/test leakage.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage framework for estimating whether rocks on rough planetary terrain can be pushed by a mobile manipulator. The visual stage segments the scene point cloud using cosine-similarity-based surface normals and DBSCAN clustering, extracts volume, an ellipsoid-fit shape error, and surface normals, and filters obstacles through a hand-designed likelihood rule. The second stage uses a Bayesian linear regressor to predict the maximum force-feedback magnitude during a push and classifies an obstacle as pushable when the predicted maximum is below the robot's capability δ (Algorithm 1). The authors report simulation experiments (90 runs, 857 recorded data points) and real-world validation of the visual preprocessing module; no quantitative evaluation of the force prediction or pushability decision is reported, and no movement ground truth is measured.","tokens_in":10962,"tokens_out":5559,"duration_ms":59943,"significance":"The application is timely and relevant: clearing easily movable rocks instead of avoiding them could improve long-term route efficiency for repeated traversals on lunar or Martian terrain. The idea of combining visual features of the rock and supporting surface with proprioceptive force feedback is sensible, and the authors state that code and data will be released. However, the evaluation does not support the central claim. There are no prediction metrics (RMSE, R², classification accuracy, confusion matrix), no validation of the force-based proxy against actual rock movement, and a data-splitting procedure that leaks information across timesteps of the same push run. The real-world experiments validate only volume and surface-normal estimation, not pushability decisions. As presented, the paper is a promising preliminary framework rather than a demonstrated system.","major_comments":[{"comment":"Pushability is defined as the probability that the obstacle moves, but the target variable learned and thresholded is the predicted maximum force magnitude f_max, and rock movement is never measured as ground truth. The decision rule in Algorithm 1 compares the predicted f_max to δ, and Section IV states a 20 N threshold that was 'empirically observed' without being derived from any measured movement outcome. A rock that moves despite a high transient force, or one that does not move despite a low force, would be misclassified, and the reported data cannot distinguish these cases. This is the load-bearing validation gap for the paper's central claim.","section":"Section III-A and Algorithm 1"},{"comment":"The dataset construction and train/test split invalidate any quantitative claim that might be drawn from the learning experiment. The paper states that 'each timestep in the point cloud ROS bag is added to the dataset along with the resulting force signal' and that 'the maximum force-feedback magnitude for each run is recorded as the label for all data points in that run'; the dataset is then 'shuffled and split into training and test sets.' Consequently, multiple timesteps from the same push run, which share one force label, fall into both training and test sets, producing optimistic estimates. No accuracy, RMSE, R², or classification metrics are reported anywhere, so the actual effect of this leakage cannot be assessed.","section":"Section IV, dataset construction and split"},{"comment":"The real-world experiments only validate the visual preliminary prediction module: estimated rock volumes (Figure 8) and average surface normals (Figure 3 and surrounding text). There is no real-world force-feedback prediction and no real-world pushability decision. The conclusion's claims about pushability therefore rest entirely on the simulation experiments, which are subject to the proxy and leakage concerns above.","section":"Section IV, real-world experiments"},{"comment":"The 'Lessons Learned' section lists possible contact outcomes—translational movement, rotational movement, slipping of the end effector, or no movement—and states that these depend on contact pose, orientation, local rock surface shape, and normal vector. This directly undermines the assumption that a single scalar f_max is a sufficient statistic for pushability, and no experimental evidence is provided to quantify how often each outcome occurs or how well f_max discriminates them. The central mapping from predicted force to pushability is therefore unverified even in the simulation domain.","section":"Section V-B"}],"minor_comments":[{"comment":"The text says '0.1π which is roughly equal to 5 degrees,' but 0.1π radians is approximately 18 degrees, not 5 degrees. This numerical error should be corrected, and the reported slope angle should be reconciled with the stated 8-degree test ramp.","section":"Section IV, real-world surface normal text"},{"comment":"The pseudocode's Bayesian update is not written consistently: the evidence term p(F_peak,i | R_i) is not the proper normalizer for a standard posterior update, and the procedure does not match the statement that Scikit-learn's BayesianRidge is used for fitting. The pseudocode should be corrected or replaced with the actual inference procedure.","section":"Algorithm 1"},{"comment":"The definition of the supporting surface S_j uses the condition z_min ≤ p_z, which appears to include all ground points above the lower z-bound of the bounding box rather than only points inside the scaled 3D box. The geometric condition should be made precise.","section":"Section III-A, surface definition"},{"comment":"There are several typographical and formatting issues, such as 'Figure 2.' with a stray period before the colon, broken equation text ('p i ∈R 3.wheren is'), and inconsistent notation for the DBSCAN output (e.g., the set notation with 'u' is unclear). These should be cleaned up in revision.","section":"General presentation"},{"comment":"The heuristic likelihood g_j = (δ/(V_j E_j))(1 - θ/π) contains several free design choices, and no sensitivity analysis or ablation is provided for the thresholds T_cs, T_low, T_high, or the DBSCAN parameters. Since the downstream regressor uses features derived from this stage, the robustness of the full pipeline to these choices is unclear.","section":"Section III-B.1, VPP likelihood"}],"recommendation":"reject","confidential_remarks":"This reads like a preliminary workshop report rather than a complete journal submission: the central empirical claim is unsupported, the data split is not valid for the stated learning problem, and the real-world experiments do not exercise the proposed affordance model. The GitHub link promises future release of code and data, but no artifacts are currently available for verification. If the authors rework the study with movement labels, run-level cross-validation, and quantitative metrics, a resubmission could be considered; as it stands, I cannot recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things. First, this paper tackles a genuine gap: instead of avoiding rocks on rough planetary terrain, a mobile manipulator clears the ones that are pushable. Second, the evidence does not support the central claim. Pushability is defined as the probability that a rock moves, but movement is never measured. The model predicts peak contact force, and the decision rule thresholds that prediction with a value the authors set on the same data (20 N in simulation). The paper itself notes in Lessons Learned that contacts can produce translation, rotation, slipping, or no movement, so an identical force trace from a stalled arm and from a rock that moved after static friction broke cannot be distinguished. The mapping from peak force to pushability is simply unverified.\n\nWhat is genuinely new and useful: the specific integration of normal-based segmentation (cosine similarity plus DBSCAN), visual features (volume, ellipsoid-fit error, surface normals), and Bayesian ridge regression to estimate contact force for rocks on sloped terrain. The dataset of 90 simulated pushes with point clouds and force signals is a real contribution, and the real-world VPP experiments for volume and surface-normal estimation are sensible and honestly presented. The authors are also unusually candid in their limitations section, which makes the central gap more visible rather than less.\n\nThe soft spots are serious. First, no ground-truth movement label exists, so the affordance itself is never evaluated. Second, the train/test split is done after shuffling individual point-cloud timesteps, and every timestep from the same push run shares one force label; that leaks information across the split. Third, no prediction accuracy, R-squared, or classification metrics are reported; the dependency graph shows coefficients but not predictive performance. Real-world tests only validate the visual module, not the pushability decision. The \"empirically observed\" 20 N threshold is asserted, not derived from measured outcomes, and delta is never given an operational value. These are not minor blemishes; they remove the quantitative support for the paper's main conclusion.\n\nWho is this for? Researchers working on non-prehensile manipulation for field or planetary robots. The problem and dataset are worth engaging with, but not yet as evidence. I would not cite this paper for the pushability claim in the next year. It does deserve a serious referee though: the flaws are fixable in revision (movement labels, split by push run, threshold from outcomes, or reframing the claim as peak-force prediction). Send it to review, but with a clear expectation of major revision.","headline":"A real problem and a real dataset, but the central pushability claim is unvalidated: movement is never measured, the force threshold is fit on the same data, and the split leaks.","tokens_in":11424,"tokens_out":1901,"would_cite":false,"duration_ms":23126,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"One exploratory push tells a robot whether a rock on rough terrain is pushable, by predicting the peak force it will feel and comparing that to its own strength.","keywords":["rock pushability","Bayesian linear regression","affordance learning","force feedback","planetary terrain","point cloud segmentation","nonprehensile pushing","mobile manipulation"],"falsifier":"Run a push experiment where the robot records the force signal and also tracks the rock's displacement (e.g., by differencing the point cloud before and after the push). If a substantial fraction of rocks classified as not pushable (predicted peak force above $\\delta$) still move, or rocks classified as pushable fail to move, the peak-force proxy fails. Concretely, across at least 100 pushes on varied slopes and rock sizes, compare the predicted peak force against measured displacement and check whether the decision rule's accuracy is significantly below chance.","tokens_in":10484,"feed_emoji":"🪨","tokens_out":3666,"duration_ms":39464,"temperature":0.7,"pith_summary":"This paper proposes replacing obstacle avoidance with obstacle removal on rough planetary terrain. It argues that a robot can decide whether a rock is pushable by predicting, from visual features alone, the peak force its arm will experience during a push, and then checking that predicted force against the robot's physical capabilities. The claim is that a Bayesian linear model trained on a modest dataset of simulated and real push experiments can make this prediction accurately enough to classify rocks as pushable, too large to move, or too small to bother with. If true, fleets of robots could clear paths once instead of detouring around the same rocks repeatedly, improving long-term navigation efficiency in places like lunar or Martian surfaces.","feed_headline":"One push tells a robot if a rock is clearable","feed_subtitle":"Bayesian model predicts peak push force from rock size, shape, and slope to clear paths instead of avoiding rocks.","key_machinery":"The central mechanism is a Bayesian linear regressor (implemented with BayesianRidge) that predicts the maximum force-feedback magnitude $f^j_{\\mathrm{max}}$ of a push action, conditioned on features $R_i = (V_i, E_i, N_i, g_i)$: obstacle volume, ellipsoid-fit shape error, surface normal set, and a visual likelihood score. The training phase initializes a prior $\\mathbf{w} := \\mathcal{N}(0, \\lambda^{-1} I)$ and updates it with Bayes' rule on observed force feedback; the prediction phase computes the posterior expected peak force and applies the decision rule $\\hat{F}_{\\mathrm{peak},j} \\le \\delta \\Rightarrow$ Pushable. This probabilistic treatment provides uncertainty quantification and lets the visual preliminary estimate $g_j$ serve as a prior belief, which the proprioceptive interaction then refines.","core_discovery":"The paper's central claim is that pushability affordance—whether an obstacle will move when pushed—can be estimated as the probability distribution of the maximum force feedback during a push, conditioned on visually observable features: the rock's volume, its shape (quantified by ellipsoid-fitting residual error), and the surface normals of the ground it rests on. The authors build a pipeline that segments rocks from point clouds using surface-normal cosine similarity and DBSCAN clustering, fits 3D ellipsoids to score shape, estimates surface normals relative to the robot, and feeds these features into a Bayesian linear regressor trained on force-feedback labels from 90 simulated and real push experiments. The regressor predicts the peak force magnitude $\\hat{F}_{\\mathrm{peak},j}$, and Algorithm 1 classifies the rock as pushable when that prediction is at or below the robot's capability $\\delta$. The paper reports that the model's predictions are strongly driven by rock volume and surface normal, and insensitive to the rock's position in the point cloud, implying the robot can assess pushability at any timestep before committing to a push.","pith_inferences":["A natural testable extension is to replace the peak-force proxy with actual rock displacement measured by differencing point clouds before and after the push; this would provide the ground-truth label the current framework never measures and could retrain the model to be robust to rocks that move despite high resistance.","The same Bayesian conditioning could be extended to predict not just the peak force but the full force-time profile, which would allow the robot to distinguish rocks that yield after a sticking point from those that never move, a distinction the peak value alone cannot capture.","The paper's own lessons-learned note that contact with rough concave/convex rock surfaces causes rotation, slipping, or no movement depending on local contact pose; a more detailed local contact-surface feature at the push point could improve prediction and is an implicit next step.","The framework's logic could be transferred to other actuators, such as the quadruped's legs for kicking or jumping, by learning a mapping from the same visual features to the different force profiles those actions produce, which the authors list as future work."],"forward_implications":["If the prediction is accurate, a robot can clear pushable rocks from a path instead of avoiding them, reducing the detour time for itself and for all subsequent agents traversing the same route.","The visual preliminary filter can pre-sort obstacles into Static (too large), Pushable, or Override (too small to bother), so the robot only wastes a push on rocks that have a real chance of moving.","Because the model is Bayesian, the robot gets a distribution over peak force rather than a single number, enabling risk-aware decisions: it can decline to push when the uncertainty crosses a safety threshold.","The finding that surface normals and volume dominate the prediction, while rock position does not, suggests the robot can evaluate pushability from any viewpoint, which could simplify the approach strategy.","The framework is intended to generalize across simulated and real rough terrain, with the real-world validation showing accurate volume and slope estimation even when the test ramp slope (8 degrees) differs from the model's estimate (5 degrees)."],"supporting_citations":[{"why":"Defines the concept of affordances in robotics that the paper operationalizes as pushability.","marker":"[17]"},{"why":"Supplies the rock shape analysis method (ellipse fitting) that the paper extends from 2D to 3D ellipsoid fitting for shape scoring.","marker":"[8]"},{"why":"Motivates the use of active haptic exploration, i.e., performing a push to gain information beyond visual cues.","marker":"[20]"},{"why":"Provides a prior method for pushing with force feedback that the paper's proprioceptive guidance builds on.","marker":"[12]"},{"why":"Demonstrates goal-driven robotic pushing using tactile and proprioceptive feedback, a baseline for the affordance learning approach.","marker":"[16]"},{"why":"Shows push effect prediction with object-centric representations, related to the paper's aim of predicting push outcomes.","marker":"[42]"},{"why":"Establishes the use of proprioceptive sensing on planetary rovers, supporting the integration of force feedback into navigation.","marker":"[34]"}],"fun_headline_variants":["Learn rock pushability with visual and force cues","Robot predicts if a rock yields to a shove","Clear rocks, don't reroute: push affordance learned","Bayesian model sizes up rock push force from slope"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system never measures whether a rock actually moves; it assumes the peak force felt during a push is a reliable indicator of pushability, so a rock that moves despite high resistance, or resists lightly yet stays put, would break the classification.","fun_headline_variants_meta":{"raw":{"variants":["Learn rock pushability with visual and force cues","Robot predicts if a rock yields to a shove","Clear rocks, don't reroute: push affordance learned","Bayesian model sizes up rock push force from slope"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1708,"prompt_tokens":978,"completion_tokens":730,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":666}},"tokens_in":594,"tokens_out":730,"duration_ms":8050,"temperature":1.0,"reasoning_tokens":666,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:23:15.628195+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a push experiment where the robot records the force signal and also tracks the rock's displacement (e.g., by differencing the point cloud before and after the push). If a substantial fraction of rocks classified as not pushable (predicted peak force above $\\delta$) still move, or rocks classified as pushable fail to move, the peak-force proxy fails. Concretely, across at least 100 pushes on varied slopes and rock sizes, compare the predicted peak force against measured displacement and check whether the decision rule's accuracy is significantly below chance.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the concept of affordances in robotics that the paper operationalizes as pushability."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the rock shape analysis method (ellipse fitting) that the paper extends from 2D to 3D ellipsoid fitting for shape scoring."},{"cited_title":"J., Valyear, K","cited_arxiv_id":null,"evidence_quote":"Motivates the use of active haptic exploration, i.e., performing a push to gain information beyond visual cues."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides a prior method for pushing with force feedback that the paper's proprioceptive guidance builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates goal-driven robotic pushing using tactile and proprioceptive feedback, a baseline for the affordance learning approach."},{"cited_title":"E., Erdem, A., Erdem, E., Asfour, T., & Ugur, E","cited_arxiv_id":null,"evidence_quote":"Shows push effect prediction with object-centric representations, related to the paper's aim of predicting push outcomes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the use of proprioceptive sensing on planetary rovers, supporting the integration of force feedback into navigation."}],"review_version":1}